How to defend against attacks in Federated Partial Label Learning (FedPLL) is a brand new and challenging question in machine learning security due to stealthy and efficient attack behaviors of adversaries. In this paper, we systematically study this problem by developing an Adaptive Partial Label Attack (APLA) which subtly manipulates the candidate label set of the data sample. To defend against APLA, we develop the RobustFedPLL framework incorporating three modules: (1) in preliminary clustering, we implement a Gaussian Mixture Model (GMM) and a moving average mechanism to identify clients' confidence; (2) in representation contrasting, we develop a contrast-based algorithm to obtain clients' model feature representations; (3) in final clustering, we utilize mainstream clustering algorithms to finally distinguish adversaries. Experiments of RobustFedPLL and SOTA defense algorithms based on two datasets are conducted, demonstrating the superiority of RobustFedPLL under various experimental settings.

Online data processing platforms offer free-tier services to attract users, where the service provider must strategically utilize limited resources to handle large amounts of low-value requests. Admission control has been leveraged to choose requests to process or skip, yet request uncertainties like unknown rewards and request numbers, and complicated interdependence among multi-resource consumption further pose challenges. To tackle these challenges, we propose a novel admission control solution, the Online Multi-resource Magician's Admission (OMMA) algorithm, that balances resource consumption with reward accumulation across online requests, while coordinating the intertwined consumption of different resources. OMMA is an online algorithm, with its performance evaluated by competitive ratio. OMMA attains a competitive ratio of C(1-K-1/2)M(1-L-1/2)N, where C is the constant capturing the intrinsic limitation of resource availability, K is the individual resource budget, M is the number of individual resources, L is the joint resource budget, and N is the number of joint resources. The competitive ratio achieved by OMMA is tight, meaning that it can be achieved by OMMA in certain problem instances. We implement trace-driven experiments to evaluate the practical performance of OMMA on a real-world LLM prompt dataset, demonstrating the superior performance of OMMA in online data processing services.

Federated unlearning (FU) algorithms offer participants in federated learning (FL) the "right to be forgotten'' for their individual data and its impact on a collaboratively trained model. Existing FU algorithms primarily focus on accelerating the retraining process and enhancing the utility of the retrained models following data removal requests. However, these approaches generally lack consideration for the robustness of FU algorithms in potential adversarial environments, where adversaries can craft malicious data removal requests to compromise the retrained model. In this work, we introduce a robust federated unlearning framework (robustFU) which notably enhances the resilience of FU algorithms against a wide range of adversarial attacks. In robustFU, we design a novel dynamic conflict sample compensation algorithm that dynamically reintroduces randomly generated samples with significant information gain to the participating clients during retraining. Additionally, robustFU employs an innovative global reweighting mechanism which adjusts the weight of each model update during the global aggregation, based on its degree of misalignment with the trained model prior to unlearning. Extensive experiments demonstrates the effectiveness and robustness of the proposed robustFU framework under adversarial environments. Furthermore, robustFU significantly accelerates the retraining process, achieving a 2.53× speed-up compared to the retrain from the scratch baseline.

Recent advances in Deep Neural Networks (DNNs) have dramatically improved the accuracy of DNN inference, but also introduce larger latency. In this paper, we investigate how to utilize early exit, a novel method that allows inference to exit at earlier exit points at the cost of an acceptable amount of accuracy. Scheduling the optimal exit point on a per-instance basis is challenging because the realized performance (i.e., confidence and latency) of each exit point is random and the statistics vary in different scenarios. Moreover, the performance has dependencies among the exit points, further complicating the problem. Therefore, the optimal exit scheduling decision cannot be known in advance but should be learned in an online fashion. To this end, we propose Dynamic Early Exit (DEE), a real-time online learning algorithm based on contextual bandit analysis. DEE observes the performance at each exit point as context and decides whether to exit or keep processing. Unlike standard contextual bandit analyses, the rewards of the decisions in our problem are temporally dependent. Furthermore, the performances of the earlier exit points are inevitably explored more compared to the later ones, which poses an unbalance exploration-exploitation trade-off. DEE addresses the aforementioned challenges, where its regret per inference asymptotically approaches zero. We compare DEE with four benchmark schemes in the real-world experiment. The experiment result shows that DEE can improve the overall performance by up to 98.1% compared to the best benchmark scheme.